How AI Can Help the Healthcare Workforce Crisis Without Replacing Clinicians

By Harvey Castro, MD, MBA

The framing of this debate has been wrong from the beginning.

The question is never whether a machine can do a clinician's job. The question is how much of a clinician's day is not clinical work at all, and how much of that can be given back.

Start with the arithmetic, because the arithmetic is what makes this urgent rather than interesting.

What workforce projections can tell us

The AAMC's March 2024 report projected a national shortage of up to 86,000 physicians by 2036. Its 2021-to-2036 projections included population growth of 34.1 percent among people aged 65 and older and 54.7 percent among those aged 75 and older. The report described 20 percent of the clinical physician workforce as aged 65 or older and another 22 percent as aged 55 to 64. Those figures belong to that report's baseline, not a new 2026 workforce measurement (AAMC March 21, 2024 report announcement).

Nursing follows a similar curve with a sharper rural edge. HRSA projects a shortfall of 108,960 full-time registered nurses by 2038, with demand met at 97 percent nationally but only 89 percent in nonmetropolitan areas. The licensed practical nurse picture is worse: a projected shortfall of 245,950 FTEs, with just 70 percent of demand met by 2038 (HRSA Bureau of Health Workforce).

Training pipelines take a decade. Immigration policy moves slowly. Compensation competition moves clinicians between systems without creating any new ones.

So the supply side is largely determined. What remains adjustable is the work itself.

Take the work nobody trained for

Ambulatory clinicians spend more time on documentation and administrative tasks than on direct patient care. Nobody applied to medical school for that.

In a voluntary, uncontrolled 30-day pre/post quality-improvement study of 263 ambulatory clinicians across six health systems, self-reported after-hours documentation decreased by a mean of 0.90 hours per week. Adjusted burnout estimates decreased from 51.9 percent to 38.8 percent in the smaller burnout-analysis sample. These findings describe an association and do not establish causation (Olson et al., JAMA Network Open, 2025). A separate three-month Stanford quality-improvement pilot enrolled 48 physicians; paired survey analyses of 38 respondents found reductions in task load and burnout (Shah et al., JAMIA, 2025).

The reported reduction was in self-reported after-hours documentation time per week. This short-term, uncontrolled study does not establish a daily staffing-capacity gain or a return on investment for another organization.

The adoption curve is already moving. Among the 2,784 US hospitals on Epic, 62.6 percent had adopted ambient AI documentation tools as of June 2025 (Yang F, Graetz I, American Journal of Managed Care, 2026).

Expand the reach of scarce specialists

The second lever is reading volume in specialties where the pipeline is thinnest.

The MASAI trial randomized 105,934 women in the Swedish national screening program to AI-supported screening or standard double reading. AI-supported screening found 6.4 cancers per 1,000 participants against 5.0 per 1,000, a 29 percent increase, with more small node-negative invasive cancers detected and no significant rise in false positives. Screen-reading workload fell 44.2 percent. According to PubMed, that is Hernström V, Josefsson V, Sartor H, et al., Lancet Digital Health, 2025 (DOI).

Read that result carefully, because the design is the lesson. The AI triaged which examinations needed one reader versus two. The radiologist still read and still decided. More cancers found, fewer reading hours consumed, no radiologist replaced.

That is the shape of a workforce intervention worth funding.

The third lever, and the one nobody funds

Most workforce conversations stop at clinicians. The bigger staffing hole in many hospitals is everything around them.

Scheduling. Prior authorization. Referral coordination. Discharge instructions written at a reading level patients can use. Translation. Chart abstraction for quality reporting. Those roles are hard to fill, cheap to burn out, and almost entirely made of structured administrative work, which is exactly what current systems handle well.

I feel this one personally. I practice in an emergency department in Texas and I speak Spanish natively. The number of times a discharge conversation has been rushed because no interpreter was available at 1 a.m. is not small. Language support is a workforce problem dressed up as a technology problem, and it is one of the few places where a tool can raise quality and capacity in the same move, provided a human verifies anything clinically consequential.

Fund that layer. It is less exciting than diagnostic AI, it faces a much lower safety bar, and it protects the clinical workforce by keeping non-clinical work off their desks.

Where AI must not go

I want to be equally direct about the boundary, because the credibility of everything above depends on it.

These laws cover different decisions and actors.

Indiana's HB 1271 added a medical-necessity downcoding provision effective July 1, 2026: an insurer may not use automation, including AI, as the sole basis without an employee or contractor reviewing the covered person's medical record. That clause does not specify a licensed clinician, and the chapter excludes Medicaid and Medicaid managed-care organizations.

Utah's SB 319, effective January 1, 2027, addresses adverse preauthorization determinations involving clinical or medical necessity. The decision-maker must know the condition or consult a knowledgeable specialist, exercise independent medical judgment subject to the specified specialist-consultation exception, and not rely solely on another source's recommendations. Separate provisions govern licensed review of specified appeals.

Georgia's SB 444, effective January 1, 2027, requires a qualifying natural person to conduct utilization review with a clinical peer participating before AI or software issues an adverse determination; the software cannot supersede that peer's judgment.

Washington's prior-authorization statute also limits AI-only adverse decisions and requires licensed review for medical-necessity denials within its defined carrier and benefit-manager scope.

Delaware's HB 191, effective April 23, 2026, bars nonhuman entities from specified nursing, physician, and physician-assistant licenses or certification and from using the corresponding protected titles.

Behavioral health has drawn the hardest lines. Maine's HB 2082 restricts mental health professionals to administrative uses of AI. Arizona requires documented informed consent before AI is involved in behavioral health services beginning January 1, 2027. Idaho and Nebraska passed Conversational AI Safety Acts effective July 1, 2027 requiring chatbots to disclose that they are AI and to implement suicide-response protocols.

Those boundaries are correct, and they are correct for a reason that has nothing to do with technology. A person in crisis needs a person.

The staffing failure mode to avoid

Here is how a good tool turns into a bad outcome.

A system deploys ambient documentation, clinicians get an hour back, and finance responds by increasing panel size or shortening visit length until the hour is gone. Burnout returns, trust in the tool collapses, and the next initiative meets a harder audience.

Decide in advance what the recovered time is for. Patient access, teaching, a shorter workday, or a smaller locums line. Write the decision down before go-live, because after go-live it becomes a budget negotiation and the clinicians lose that one every time.

Where the equity risk sits

The systems that most need capacity relief are the least likely to get it. Ambient AI adoption ran at 64.7 percent in metropolitan hospitals against 54.3 percent in nonmetropolitan hospitals, 70.2 percent at nonprofits against 28.8 percent at for-profit facilities, and 67.6 percent in the highest operating-margin quartile against 58.0 percent in the lowest (American Journal of Managed Care, 2026).

Now overlay HRSA's projection that nonmetropolitan RN shortages reach 11 percent by 2038 against 2 percent in metro areas. The places with the thinnest workforce are adopting the capacity tools slowest.

That gap will not close on its own. It closes through purchasing coalitions, state and regional programs, and health systems that extend enterprise agreements to their rural affiliates rather than piloting only at the flagship.

What I would tell a CEO

Fund the tools that remove work clinicians never trained to do. Fund the tools that stretch a scarce specialist's reading day. Keep licensed humans on every decision that denies, delays, or diagnoses. And decide who owns the time you free up before you free it.

AI will not replace clinicians in my lifetime. It might make it possible for enough of them to stay.

Harvey Castro, MD, MBA (DR GPT™) is a board-certified emergency physician, 5x TEDx speaker, Advisor to the CEO of Phantom Space, and author of books on AI, medicine and leadership.

Related DR GPT™ reading

• The Future of Healthspan and AI: From Sick Care to Predictive Medicine
• Healthcare AI Keynote Speaker: What Hospital Leaders Need to Hear in 2026

More from DR GPT™: who DR GPT™ is, healthcare AI keynotes, board advisory, speaking, TEDx talks, books, and the media kit.

Book Harvey Castro, MD, MBA, known as DR GPT™, for a healthcare AI keynote, board retreat, or executive session. Start the conversation.